Unsupervised Selective Rationalization with Noise Injection

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 12647–12659 July 9-14, 2023 ©2023 Association for Computational Linguistics Unsupervised Selective Rationalization with Noise Injection Adam Storek Columbia University astorek@cs.columbia.edu Melanie Subbiah Columbia University m.subbiah@columbia.edu Kathleen McKeown Columbia University kathy@cs.columbia.edu Abstract A major issue with using deep learning models in sensitive applications is that they provide no explanation for their output. To address this problem, unsupervised selective rationalization produces rationales alongside predictions by chaining two jointly-trained components, a ra- tionale generator and a predictor. Although this architecture guarantees that the prediction re- lies solely on the rationale, it does not ensure that the rationale contains a plausible explana- tion for the prediction. We introduce a novel training technique that effectively limits gen- eration of implausible rationales by injecting noise between the generator and the predictor. Furthermore, we propose a new benchmark for evaluating unsupervised selective rationaliza- tion models using movie reviews from existing datasets. We achieve sizeable improvements in rationale plausibility and task accuracy over the state-of-the-art across a variety of tasks, in- cluding our new benchmark, while maintaining or improving model faithfulness.1 1 Introduction With the advent of large pre-trained language mod- els like GPT-3 (Brown et al., 2020), the size and complexity of deep learning models used for natu- ral language processing has dramatically increased. Yet greater performance and complexity can come at the cost of interpretability, masking anything from implementation mistakes to learned bias. A model architecture that justifies its output by providing relevant subsets of input text as a ratio- nale is therefore desirable (see example in Figure 1). The unsupervised selective rationalization archi- tecture as introduced by Lei et al. (2016) generates rationales alongside predictions by chaining two jointly-trained components, a rationale-generator and a predictor. The generator extracts a rationale: concatenated short and concise spans of the input 1Code and benchmark are available at https://github. com/adamstorek/noise_injection. Class: Positive Ultra low budget but extremely inventive horror film about a group of friends vacationing in a cabin who accidentally awaken an evil force in the woods via the necronomicon, the book of the dead. Bruce Campbell stars as Ash, who eventually becomes the sole survivor and has to battle both the demons from the woods, and his friends who have become demons (including his own girlfriend). The results shown on screen are amazing considering the film's tiny budget, constant location changes, and a filming schedule that was sporadic over two years… Movie Review: Figure 1: Example of a rationale selected by BERT- A2R + NI (our model) on the USR Movie Review dataset (our benchmark), which asks models to clas- sify movie reviews as positive or negative. text that suffice for prediction. The predictor bases its prediction only on this rationale, which encour- ages faithfulness, meaning how much the rationale reveals what parts of the input were important to the model’s prediction. In practice, however, the rationale often isn’t plausible, meaning it can’t convince a human of the correct prediction, …

Content truncated to 500 words